Project 55: Extensions to the BLAS Papers published in TOMS. ---------------------------------------------------------------------- Pasadena 1984: ============== Sparse vector extensions to the BLAS (Reid) ------------------------------------------- Roger Grimes and Dave Dodson of Boeing Computer Services would like to standardize a few BLAS for sparse matrix work (see IFIP WG 2.5 (Pasadena-14) 1114). Reid thinks that we should support this, perhaps extending to a new "BLAS2" project that includes basic kernels that go beyond vector operators (e.g. to perform matrix multiplication) because this is needed for efficiency on advanced hardwares. Extending the BLAS (Hanson) --------------------------- People who may participate : Reid, Rice, Aird, Ford, Lawson, Smith (WG 2.5) Dodson, Grimes (Boeing Computer Services) Dongarra, Moler (Eispack 8x) Braams (Garching) Krogh, Kincaid (Former BLAS authors) Plan . 1. Propose to ACM-SIGNUM a joint WG 2.5 ACM-SIGNUM project: "BLASX". 2. If ACM says 'yes' (likely!) invite key people to contribute ideas for a minimal set of names. 3. Schedule open meeting(s) with draft of results of 2 when available. 4. Modify draft; publish in ACM-SIGNUM Newsletter. 5. Implement results in a model version: Fortran subroutines. 6. Write test driver for new routines. 7. Publish algorithm consisting of 5, 6. Sophia-Antipolis 1985: ====================== BLAS2 (Hanson) -------------- This work intends to describe an extended set of Basic Linear Algebra Subprograms (in Fortran). The extensions are targeted at matrix-vector operations which provide for more efficient and portable implementations of algorithms for high-performance computers. The WG 2.5 is only tangentially involved in this project. A personal note - this is a good example of the WG 2.5 pointing out an area that needs action, but non-WG 2.5 people actually do most of the work. The proposal has been printed in ACM-SIGNUM Newsletter, 1985. Some revision of the proposal is expected based on comments from the user community. A further proposal, unrelated, has been made for the set of sparse vector BLAS. This document, printed in the ACM-SIGNUM newsletter, was circulated at the 1984 WG 2.5 meeting in a somewhat different farm. This work has no official connection with WG 2.5. The group welcomed both pieces of work on extending the BLAS. Argonne 1986: ============= BLAS 2 (Extended BLAS) ---------------------- R. Hanson presented a brief history of the BLAS project and described the current state of its development. During the years 1984-1986 the level 2 (or extended BLAS) were submitted to informal review from the Applied Math community. Based on this review a few changes were made to initial specifications of the package as presented in the January 1985 issue of ACM SIGNUM Newsletter. The authors of the package felt that the the specification was now firm. A testing procedure and accompanying Fortran 77 package were developed. This test program will accompany the current Fortran 77 model implementations. Points still for discussion are precision and testing of BLAS 2. Implementations tailored for particular machines are expected and will be encouraged in the published documentation. In the ensuing discussion it was noted that there was a certain amount of overlap between BLAS and Fortran 8X elementary functions. However, names and call sequences are different which could be a possible source of conflicts (Paul). Routine names should become more intelligible and coordinated with Fortran 8X (Reid). There are currently about 16 real and double precision routines, and about 70 complex routines in BLAS 2. Hanson will handle any written suggestions concerning BLAS2. A paper on BLAS 2 is to be published in SIGNUM. B. Smith will send copies of the paper to all WG 2.5 members that requested it. Motion by B. Ford (seconded by Lawson): WG 2.5 thanks R. Hanson on the effort he put into the BLAS project. Como 1987: ========== Ford talked about planned extensions of BLAS noting that they dealt primarily with order n^3 operations and are intended for the blocks of partitioned matrices. A discussion developed. It was noted that the BLAS 3 should be able to make full use of machine architecture including distributed memory machines (Delves, Smith, Rice, Gentleman, Stetter). Documents: ========== 1. J. J. Dongarra, J. Du Croz, S. Hammarling, and R.J. Hanson, "An Extended Set of FORTRAN Basic Linear Algebra Subprograms," ACM Trans. Math. Soft, Vol. 14 (1), 1-17, 1988. 2. J. J. Dongarra, J. Du Croz, S. Hammarling, and R.J. Hanson, "ALGORITHM 656 - An Extended Set of Basic Linear Algebra Subprograms: Model Implementation and Test Programs," ACM Trans. Math. Soft, Vol.14 (1), 18-32, 1988. 3. J. Dongarra, J. Du Croz, I. Duff, S. Hammarling, "A Set of Level 3 Basic Linear Algebra Subprograms", ACM Trans. Math. Soft., Vol.16 (1), pp 1-17, 1990. 4. X. Li, J. Demmel, D. Bailey, G. Henry, Y. Hida, J. Iskandar, W. Kahan, S. Kang, A. Kapur, M. Martin, B. Thompson, T. Tung, D. Yoo, "Design, Implementation and Testing of Extended and Mixed Precision BLAS", ACM Trans. Math. Soft., Vol. 28 (2), pp 152-205, 2002.